Brad DeLong does something rare in the age of artificial intelligence: he treats a large language model not as an oracle or a threat, but as a surprisingly effective, if uncanny, pedagogical partner. By engaging a machine to translate a notoriously difficult passage from Cicero's In Catilinam, DeLong reveals that the "magic" of these systems lies in their ability to compress centuries of human teaching into probabilistic patterns, effectively thawing out frozen history to interact with us in real time.
The Stutter as Style
DeLong begins by wrestling with a specific syntactic challenge: a sentence from Cicero where the rhythm is intentionally heavy, filled with nested negative clauses. He notes that an early attempt by the AI model "momentarily misses" the context, creating a "strange hiccup of the attention mechanism." Yet, he quickly pivots to what actually matters in this exchange. The model offers three distinct translation strategies: one prioritizing modern clarity, another preserving literary grandeur, and a third attempting literal fidelity.
"The 'stuttering' rhythm you are describing is a hallmark of periodicity — the hallmark of Latin oratory."
DeLong argues that what looks like clunky English to a modern reader is actually a faithful reproduction of Ciceronian architecture. He explains that Cicero's power came from suspending meaning through layers of subordinate clauses, only resolving them with a final, heavy verb. The AI, by mimicking this "stutter," inadvertently captures the incantatory nature of ancient rhetoric. This is a crucial insight: the model isn't just predicting words; it is simulating a specific historical mode of thought.
However, one might ask if this simulation risks flattening the nuance of human interpretation into a mere statistical average. While DeLong celebrates the "grooves" of the training data, critics could argue that relying on these patterns reinforces established biases in how we teach classics, potentially sidelining newer, more disruptive scholarly perspectives.
"If you are writing a modern news report, your version is 'bad' English. If you are writing the speech of a Roman Senator or a high-stakes protagonist in a political thriller, your version is brilliant."
The Ghosts in the Machine
The core of DeLong's argument shifts from linguistics to epistemology. He observes that the AI works so well because it has been trained on "centuries of grammarians and teachers" who performed similar tasks for centuries. The model isn't thinking; it is re-enacting a script written by human educators long dead.
"It feels so close to having a personal grammarian on the log because the task structure is so well aligned."
DeLong draws a parallel between his own experience at Oxford and this digital interaction. He recalls how he and colleague Kevin O'Rourke once reinforced each other's points, only to realize later that they were both channeling their shared mentor, Jeffrey Williamson. "It was his mind in the seminar that was doing the thinking behind our words," DeLong writes. In this light, human learning itself is a form of "stochastic parrotry"—the acquisition of stable repertoires of moves that we redeploy without always knowing their origins.
This comparison is provocative. It suggests that the distinction between human understanding and machine generation is blurrier than we admit. If human expertise in well-institutionalized fields is largely about mastering established patterns, then a model trained on those same patterns can perform the role with "locally indistinguishable" success.
"The pedagogical power is coming from the history of Latin teaching frozen in text. And the LLM thaws it out again."
Yet, this raises a critical question about agency. If we are merely re-enacting training data, where does genuine insight come from? DeLong acknowledges that the AI's parrotry is "cruder," but he insists it nudges him into "good Latinist behavior" because both he and the machine share the same textual heritage. The system works not because it understands Cicero, but because it has seen the conversation tens of thousands of times before.
The Bottom Line
DeLong's most compelling contribution is reframing large language models as a form of cultural resurrection rather than just a computational tool; they allow us to access the "sedimented" wisdom of past pedagogical practices in an interactive format. However, his argument leaves unresolved the tension between efficient pattern recognition and the messy, non-linear nature of true intellectual breakthrough. The reader should watch for how this "thawing" of historical data might reshape not just language learning, but the very definition of expertise in any field where tradition dominates discourse.